MAS.632 Conversational Computer Systems - MIT OpenCourseWare

PROSODY
Torwd More Ronlosarmmunniikoi
measures. Under this approach, a grammar is augmented with word-sequence
probabilities based on analysis of a corpus of utterances spoken by naive subjects
attempting to access the chosen task [Hirschman et al. 1991].
But in natural speech these probabilities are dynamic and depend on the current
state of the conversation as well as the listener's expectations of what the
talker may be up to. In a fully integrated conversational application, pragmatic
information such as identification of a partially completed plan or detection of
elements of a script could tune word probabilities based on expectations of what
is likely to follow. Similarly a focus model could suggest heightened probabilities
of words relevant to attributes of entities recently discussed in addition to resolving
anaphoric references. Domain knowledge plays a role as well, from simplistic
awareness of the number of digits in telephone numbers to knowledge of the
acceleration and turning capabilities of various types of aircraft in an air traffic
control scenario.
Part of the difficulty with flexible parsing is the excessive degree of isolation
between the application, its discourse system, and the speech recognition component.
For many current systems the recognizer is given a language model in
whatever form it requires; it then listens to speech and returns a string of text to
the application, which must parse it again to know how to interpret the words
meaningfully. Not only has syntactic information been needlessly lost when
reporting the recognized speech as a string of words, but it also may be detrimental
to strip the representation of any remaining acoustic evidence such as the
recognizer's degree of certainty or possible alternate choices of recognition
results. How can partial recognition results be reported to the parser? Perhaps
several noun phrases were identified but the verb was not, confounding classification
of the nouns into the possible roles that might be expressed in a framedbased
representation.
These observations are intended to suggest that despite this book's portrayal of
the various layers of language understanding and generation as distinct entities,
they still must be tightly woven into a coherent whole. Isolation ofthe word identification
portion of discourse understanding into a well bounded "speech recognizer"
component cannot in the long run support sophisticated conversational
systems. Knowledge must be communicated easily across components, and analysis
must be flexible and based on dynamic conversation constraints.
Prosody refers to the spoken style of discourse independent oflexical content, and
it includes several aspects of how we speak. Intonation is the tune of an utterance:
how we modulate FO to change the pitch of our speech. Intonation operates
at a sentence or phrase level; the rising tune of a yes-or-no question immediately
differentiates it from the falling tune of a declarative statement. Intonation also
helps to convey the stress of words and syllables within a sentence as stressed
syllables are spoken with a pitch higher or lower than normal, and specific words
are emphasized-an important aspect of communicating intent-by stressing
301